Creativity
Ben Steyn
Creativity is the disposition to produce outputs that are original and valuable.
Do AI systems exhibit creativity in their art, music, and words?
Key Points:
AI systems exhibit weak creativity: they produce original, valuable, and surprising outputs.
AI systems lack strong creativity to the extent that they are not agents.
AI systems can produce art, music, and words to a high aesthetic standard. Consider Sony Computer Science Laboratories’ DeepBach, which has fooled discerning Bach fans that they were listening to compositions by the man himself. But is AI in fact creative, or is this just an illusion?
This question matters. For some, a creative AI is unsettling, as creative acts have historically been described as the hallmark of humanity. For others, particularly those working in the creative industries, considerations about artificial creativity have immediate economic implications. Creativity is an operable term in intellectual property law, for instance, in cases around AI mimicking a human artist’s style. AI systems seem poised to replace humans in some creative work, such as in aspects of film studio animation. Creativity also occurs outside so-called creative industries, for instance, in science and engineering, where AI is also making advances. The apparent creativity of AI systems may also lead to unexpected risks in their behaviour and our ability to regulate them. In all of these discussions, it is prudent to be clear in what senses AI systems can or cannot be creative.
Outputs v.s Processes
When discussing whether an AI has some faculty X (understanding, comprehension of meaning, empathy, etc.), we can often distinguish between:
AI’s outputs seeming to imply X
AI actually having the right sort of processes ‘under the hood’ for us to truly grant X
The outputs v.s. process distinction is seen in debates around machine intelligence and understanding. The Turing Test, ‘can a chatbot’s responses fool a human that it is human?’, can be passed only by considering the chathe chatbot's outputs (I.e. its responses). Meanwhile, the ‘Chinese Room Argument’ gives reason to think that we cannot infer understanding on the basis of outputs alone (see meaning). Our discussion of creativity proceeds along similar lines, first considering a weak form of creativity defined in terms of creative outputs, before turning to a stronger form, focussed on processes.
Weak (Output) Creativity
Minimally, psychologists and cognitive scientists suggest that creative acts require originality and value. A nonsensical string of words might be highly original, but not valuable, and thus not creative. Conversely, a home cook following a recipe may produce food that is tasty and nutritious, and hence valuable, but they have merely followed set instructions, and are thus unoriginal (note, however, the chef who wrote the recipe may well have been original).
Image generation AI systems like Midjourney, DALL-E, and Stable Diffusion, and text generating systems like Chat-GPT and Gemini meet this minimal definition. They produce outputs that are valuable, be that for their aesthetic properties, or because they serve some function well. Regarding their originality, while these systems train on human works and can mimic human styles (as humans do too, of course), they do not strictly duplicate work. The extent of their originality can be debated, but here we must conclude that there is at least some, and hence, weak creativity.
In the human case, philosophers have questioned whether creativity is more than just novelty and originality. Another idea proposed is that creative outputs are surprising in some way, that is, they deviate from a common language or system of thought. This surprisingness is conceptually distinct from mere originality: consider that infinite pop songs could be written, all original, while all sharing a common set of chord structures, such that none are surprising to a musically informed listener. Meanwhile, a piece of experimental jazz surprises the same listener by deviating from the common language.
A demand for surprisingness might rule out some generative AI outputs, such as those merely repeating patterns in a limited training dataset of a single artist's works. However, we can also see AI outputs can indeed surprise us, for instance, the chess community were surprised by moves made by DeepMind’s AlphaZero, which seemed to frequently defy common human-derived heuristics and strategic concepts like “rooks are worth more than bishops” and “pawns should take towards the centre” (see also AlphaGo’s famous ‘move 37’ v.s. Lee Sedol).
In summary, we could think of AI models like DeepBach, AlphaZero, and Midjourney as having already solved the equivalent of a ‘Turing Test’ for creativity: they have fooled, surprised, and pleased unwitting human audiences. But without considering the process these models undertake, perhaps this isn’t enough to fully grant creativity.
Strong (Process) Creativity
In humans, cognitive scientists typically conceive the creative process as including various phases: preparation, idea generation or search, insight (the ‘eureka moment’), then evaluation and internal scrutiny of an idea leading to a final output. One requirement of creative processes often stipulated is that the idea generation and insight phases involve spontaneity; these processes cannot follow a mechanical procedure, algorithm or rule, such that the outcome will be pre-determined at the start. AI systems arguably have comparable processes to all of the phases listed. Moreover, they seem to have spontaneity; they do not always follow a strictly deterministic mechanical procedure. For instance, note the heavy use of randomness in the training of artificial neural networks.
But even granting that AI follows process steps comparable to humans, there is a more fundamental challenge here about the process involved in AI outputs. Consider a beautiful and rare mountaintop. This natural structure might be highly original and valuable, but we wouldn’t typically call it ‘creative’. That is because creativity requires a ‘creator’. Notions of a divine creator aside, we might figuratively describe nature, or ‘mother earth’, as having been this mountaintop’s creator. But it seems that what we really need here is some sort of agent to be present, a thinker who is intentionally cultivating the mountaintop, to regard it as a creative process. Thus, AI creativity may hinge on whether it can be deemed an agent (see Agency). And relatedly, whether it has intentions and motivations of its own (see Intention). If AI lacks these properties, then we might be better calling AI a tool in some other agent’s creative process.
Moreover, for creativity, it may not be sufficient that something has motivations and intentions per se, but also that they have the right sort of motivations and intentions. In the case of an AI artwork depicting a war, we might think creativity demands that the AI understands the historical context it is rendering; why it is important to create art about this topic; and how its artistic decisions will elicit certain emotions or aesthetic sensations in a viewer.
That said, we should be careful not to be too prescriptive. Consider that there are many human examples of creativity where it seems the intentions and motivations of the humans involved were quite prosaic, for instance, scientists seeking profit or fame, not necessarily new knowledge for its own sake. Perhaps this is no different to the sense in which an AI model is ‘motivated’ by its objective function. Nevertheless, serious questions about AI’s ability to hold agency and intentions prevent us from concluding that AI has creativity in the strong sense, yet.
Other Issues
The Value of Creativity
Human creativity is considered a virtue or intrinsic value by many. The term is typically used with a positive valence – it’s a compliment. One counter-argument to this is the problem of ‘dark creativity’, for instance, an intricately plotted jewellery theft. Was the thief’s creativity to be admired? There are three options here: 1) deny this is truly creative, 2) accept this is creative but deny creativity is virtuous, or 3) ‘bite the bullet’ and accept that the thief was, in some but not all ways, doing something virtuous. Given we can’t easily settle the matter of whether creativity is valuable in the ordinary human case, it is reasonable to be unsure about the value of AI creativity (or lack thereof).
Ethics of AI Creativity
AI creativity raises practical issues around intellectual property and the devaluing of human artistic skills. Assuming human creativity is valuable, there may also be objections to the promulgation of AI creative outputs, perhaps because they are inauthentic, do not contain within them a coherent authorial perspective, they deny humans the opportunity to be creative, or are, for some inarticulable reason, not as aesthetically good as human equivalents. Meanwhile, in mixed cases of human and AI combining to create an output, there is a difficult question of how we attribute credit or ownership between them, and to what extent it is possible to value the human contribution while disvaluing the AI's.